AxForge
Sign in Get started

Model reference · open weights

RealVideo

LLMs zai-org Omni (any→any) 1 build Open weights 0 dl/mo

RealVideo is an open-weight language model from zai-org. RealVideo (BF16) weighs 65.2 GB; the smallest configuration that runs it is H100 80 GB.

  • RealVideo is a WebSocket-based video calling system developed by zai-org that supports text input and generates real-time conversational video.
  • It utilizes GLM-4.5-AirX and GLM-TTS models to produce audio responses and employs autoregressive diffusion to create corresponding video frames with lip sync.
  • The project is released under the MIT license and supports English and Chinese.

Summary of the zai-org/RealVideo model card, 2026-10-01

Hugging Face ↗GitHub ↗

What it is

Released byzai-org
Released2025-12-11
VRAM65.2 GB for the weights

What it runs on

Memory and cards for RealVideo (BF16)

65.2 GBweights, file size
762 MBruntime overhead, at least

How much memory each request adds isn't estimated yet for this architecture. The weights need at least the cards below, plus room for the context.

CardWeights alone
RTX 3060 12 GB … L40S 48 GB
6 smaller cards		does not fit
A100 80 GBfits
H100 80 GBtight
RTX PRO 6000 Blackwell 96 GBfits
DGX Spark (GB10) 128 GB unifiedfits
H200 141 GBfits
B200 180 GBfits

From the model card

What zai-org says about RealVideo

Read the model card

RealVideo is a WebSocket-based video calling system that supports text input. It leverages GLM-4.5-AirX and GLM-TTS models to generate audio responses and utilizes autoregressive diffusion to generate corresponding video frames. The system features a modular design with full functionality and a clean code structure. Visit blog here!

Features

  • Text Input: Supports text message input.
  • AI Voice Response: Integrates GLM-4.5-AirX and GLM-TTS models to generate voice responses.
  • Lip Sync: Generates real-time conversational video based on any input image and audio.
  • Real-time Communication: WebSocket-based real-time bidirectional communication.

Quick Start

you can check in our GitHub.

Technical Highlights

  • Model Integration: Allows for convenient and quick voice cloning, taking text input to generate audio output.
  • Modular Design: Clear code structure, easy to maintain and extend.
  • Real-time Performance: Optimized audio processing and real-time video generation algorithms.

Acknowledgements

This project utilizes the following open-source libraries:

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

How it works

How language models work

Your prompttext / messagesTransformerattention over tokensNext-token loopgenerate + streamResponsetext · tool callsA language model reads your tokens and predicts the next one, again and again, streaming the reply back.
© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms